🚀 go-pugleaf

In article <9149@june.cs.washington.edu> david@june.cs.washington.edu.cs.washington.edu (David Callahan) writes:
>In article <1082@cernvax.UUCP> hjm@cernvax.UUCP (Hubert Matthews) writes:
>>  Why do Crays have fast memories?
>>Because [long vector] is exactly the type of problems they are designed for.

OK, if you want me to be more specific: Why do Crays have fast memory
systems?

>[Crays use commodity RAMs]

I know, but have you seen how much stuff goes around these jellybean
RAMs so that they can provide enough bandwidth?  Lots of banks,
interleaved to get the bandwidth up.  Not the latency, but the
bandwidth.  If you do vector loads, then the latency of the first load
is offset by the speed of the following loads.  If you get the vector
stride wrong (a power of 2, for example), then your Cray's memory will
crawl along as you will be accessing one bank continuously, rather
than interleaving accesses to several banks.

>Cray's cope with long memory latencies by using vector loads,
>effectively pipelining memory. So you were right about these machines
>being designed for vector operations but really a stronger statement
>could be made: Cray machines require vector operations to acheive
>anywhere near peak performance. 

Get your interleaving wrong and you can kiss your MFLOPS goodbye
because of the memory system, which was, after all, my original point:
There is more to computing systems than just a CPU; if you don't have
the memory and the I/O to feed the CPU, then the CPU speed is wasted.
A computing system must be balanced, *with respect to the code it is
running*, if it is to be cost effective.  PCs are biased towards a
market where they offer a reasonable balance; Crays are biased towards
another sort of market.  Just don't expect me to buy a PC to crunch
a gigabyte of data or to buy a Cray to do word processing.

-- 
Hubert Matthews      ...helping make the world a quote-free zone...

hjm@cernvax.cern.ch   hjm@vxomeg.decnet.cern.ch    ...!mcvax!cernvax!hjm